Augmentation - imatge-upc.github.ioimatge-upc.github.io/.../slides/D2L2-augmentation.pdf ·...
Transcript of Augmentation - imatge-upc.github.ioimatge-upc.github.io/.../slides/D2L2-augmentation.pdf ·...
Augmentation
Day 2 Lecture 2
IntroductionImageNet Classification with Deep Convolutional Neural Networks, Krizhevsky A., 2012
ImageNet Large-Scale Visual Recognition Challenge (ILSVRC) 1.2 million training images, 50,000 validation images, and 150,000 testing images
Architecture of 5 convolutional + 3 fully connected = 60 million parameters ~ 650.000 neurons.
Overfitting!!
● Reduce network capacity
● Dropout
● Data augmentation
Ways to reduce overfitting
● Reduce network capacity
● Dropout
● Data augmentation
Ways to reduce overfitting
1% of total parameters (884K). Decrease in performance
● Reduce network capacity
● Dropout
● Data augmentation
Ways to reduce overfitting
37M, 16M, 4M parametes!! (fc6,fc7,fc8)
Ways to reduce overfitting
● Reduce network capacity
● Dropout
● Data augmentation Every forward pass, network slightly different.Reduce co-adaptation between neuronsMore robust features
More interations for convergence
Ways to reduce overfitting
● Reduce network capacity
● Dropout
● Data augmentation
Data Augmentation
During training, alterate the input image (Krizhevsky A., 2012)
- Random crops on the original image- Translations- Horitzontal reflections- Increases size of training x2048- On-the-fly augmentation
During testing
- Average prediction of image augmented by the four corner patches and the center patch + flipped image. (10 augmentations of the image)
Data Augmentation
Alternate intensities RGB channels intensities
PCA on the set of RGB pixel throughout the ImageNet training set. To each training image, add multiples of the found principal components
Object identity should be invariant to changes of illumination
Augmentation for discriminative unsupervised feature learning
Discriminative Unsupervised Feature Learning with Exemplar Convolutional Neural Networks, Dosovitskiy, A., 2014
MOTIVATION● Large datasets of training data● Local descriptors should be invariant transformations (rotation, translation, scale, etc)
WHAT THEY DO● Training a CNN to generate local representation by optimising a surrogate classification task● Task does NOT require labeled data
Augmentation for discriminative unsupervised feature learning
Select random location k and crop 32x32 window(restrictions: region must contain objects or part of the object: high amount of gradients)
Apply a transformation [translation, rotation, scalig, RGB modification, contrast modification]
...
Generate augmented dataset: 16000 classes of 150 examples each
Class k=1, with 150 examples
Augmentation for discriminative unsupervised feature learning
Generate augmented dataset: 16000 classes of 150 examples each
Example of classes Example of examples for one class
Augmentation for discriminative unsupervised feature learningClassification accuracies
Superior performance to SIFT for image matching.
Summary
Augmentation helps to prevent overfitting
It makes network invariant to certain transformations: translations, flip, etc
Can be done on-the-fly
Can be used to learn image representations when no label datasets are available.